Use this when no taskId is known to read the current project work and ordered ready queue. Choose only ready tasks with runnable=true; waiting tasks include their unfinished dependencies. Results contain task summaries only, without each task's objective or acceptance criteria; call get_task for the full definition.

Read one task status section with a bounded limit and opaque cursor. Results use the existing ready-queue or recent-completion ordering. Each task is a summary without its objective or acceptance criteria; call get_task for the full definition.

Read a known taskId's definition and the context for its current phase. The default working view returns the phase (implementation or review), dependency state, the current run's progress, every acceptance criterion with its stable id and full text, the latest submission with per-criterion delivery results, previous-round findings awaiting closure, and unclosed findings from changes-requested rounds. Pass view='full' to retrieve the complete immutable event and review histories. This does not require get_project_status first.

Record one newly discovered task in the project's inbox. Optionally include the model label. Agent-created tasks are not ready until a developer confirms and orders them. Before creating a task, confirm you can state the observable outcome, explicit exclusions, and verification. If unclear requirements prevent that, ask the developer instead of filling it in. When creating a task, check that each acceptance criterion describes one independently verifiable delivery result, has a corresponding verification, and does not smuggle in an unconfirmed implementation restriction.

Record several newly discovered inbox tasks atomically, optionally including model labels. If one input is invalid, the complete batch is rolled back. Before creating a task, confirm you can state the observable outcome, explicit exclusions, and verification. If unclear requirements prevent that, ask the developer instead of filling it in. When creating a task, check that each acceptance criterion describes one independently verifiable delivery result, has a corresponding verification, and does not smuggle in an unconfirmed implementation restriction.

Correct the definition of an inbox task created by you or another agent. Call get_task first and pass the task's latest updatedAt with the complete title, objective, acceptance criteria, and dependencyIds. The edit is rejected if the task changed since it was read, cannot change risk level or risk confirmation, and only works while the task is in inbox. Use update_task for run-scoped progress instead.

After get_task, atomically claim a runnable ready task. Optionally include the model label. Save the returned runId and include it in every subsequent execution update.

Record run-scoped progress for a running task. The runId must belong to the current execution. Use edit_inbox_task to correct an inbox task definition instead.

Submit a running task for review with a completion summary, an exact 40-character commit SHA, and exactly one delivery result per acceptance criterion, filed under the criterion id returned by get_task (met, not_met, or unverifiable; met requires recorded evidence). The task changes to in_review and the runId becomes invalid after submission.

Use this when the developer sends a task ID for independent review. Call get_task first and confirm phase=review, then atomically claim the latest pending review and save the returned reviewRunId before inspecting the submitted artifact. Review the exact commit in the latest submission. Do not accept the submitter's criteriaResults as evidence: independently reproduce the key verification for each acceptance criterion, inspect the diff for changes outside the task objective and acceptance criteria, and check the work against the repository root AGENTS.md. When a previous round requested changes, close every previous finding. Record low, medium, or high severity for each finding, but severity does not permit a pass: a review cannot pass while any submitted criterion is not_met or unverifiable, or any previous finding remains not_fixed or partially_fixed. Conclude with approve_review, request_changes, or complete_review using the reviewRunId.

Pass the claimed review with a summary, structured findings, finding closures, and verification evidence. Every new finding records issue, suggestion, severity, trigger, impact, and evidence, plus a location when one can be identified. When the previous round requested changes, record a closure status for every previous finding; unresolved or partially fixed findings and not_met or unverifiable delivery results block passing. The reviewRunId must own the current review; low-risk tasks may complete automatically while medium- and high-risk tasks wait for developer approval.

Return the claimed review for rework with actionable findings and finding closures. Findings require trigger, impact, and evidence; a re-review must close every finding from the previous changes-requested round. The reviewRunId must own the current review; the task returns to ready and requires a new execution runId.

Record a passed or changes-requested AI review with structured findings, finding closures, and verification evidence. Passing is rejected when submitted acceptance criteria are not met or unverifiable, or when a previous-round finding remains not_fixed or partially_fixed. Completion also requires the fixed review commit to be present on the developer-confirmed local target branch; low-risk tasks may complete automatically only after that check, while medium- and high-risk tasks wait for developer approval.

Retry the local read-only Git integration check for an already accepted latest review. Supply the expected review id from get_task. This never merges, fetches, pushes, or substitutes another commit SHA.

Read-only. Returns the full MileStack agent instructions, identical to the .milestack/agent-instructions.md that milestack init generates for this version. Call it when that file is missing in the current checkout.